Liwc Psycholinguistic

LIWC Psycholinguistic Analysis

Overview

LIWC (Linguistic Inquiry and Word Count) psycholinguistic analysis categorizes a corpus's vocabulary into hierarchical psychological dimensions -- cognitive processes, social dynamics, affective states, biological concerns, drives, and perceptual processes -- then computes each dimension's proportional weight to identify which linguistic registers dominate the writing. The core principle: words are windows into psychological processes, but dictionary-based categorization captures word-level signals, not meaning-level understanding. A word like "cut" matches the biological/body category regardless of whether the text describes surgery, budget cuts, or a film edit. Treat dimension percentages as statistical summaries of vocabulary composition, never as direct measures of psychological states.

Methodological foundation: LIWC-22 (Boyd et al., 2022) provides ~6,500 dictionary entries mapped to 118 hierarchically organized categories. Open-source alternatives include Empath (Fast et al., 2016; ~200 categories, r = 0.90 correlation with LIWC), SEANCE (Crossley et al., 2017; 254 indices with negation and POS features), and custom dictionary approaches using validated word lists. All share the same bag-of-words limitation: no handling of negation, sarcasm, polysemy, or compositional meaning.

When to Use

• Categorizing an author's vocabulary into psychological dimensions for style profiling
• Identifying which linguistic registers (cognitive, social, emotional, perceptual) dominate a corpus
• Establishing dimension baselines to compare across corpora, time periods, or contexts
• Mapping psycholinguistic dimension distributions to style replication priorities
• Providing input features for downstream personality, archetype, or voice modeling analyses

When NOT to use:

• Diagnosing psychological conditions, mental health states, or clinical profiles
• Corpus contains fewer than 500 words (see Insufficient Data Handling)
• Text is heavily non-English (LIWC dictionaries are English-calibrated)
• You need semantic understanding of how words are used in context (dictionary methods cannot provide this)
• Making evaluative judgments about a person's character or mental fitness
• Text is machine-generated, heavily templated, or formulaic (dimension percentages reflect the template, not a person)

digraph liwc_decision {
    rankdir=TB;
    "Corpus 500+\nwords?" [shape=diamond];
    "Primarily\nEnglish?" [shape=diamond];
    "Need semantic\ncontext?" [shape=diamond];
    "Clinical/diagnostic\npurpose?" [shape=diamond];
    "Proceed with\nLIWC analysis" [shape=box];
    "Report insufficient\ndata" [shape=box];
    "Flag language\nlimitation" [shape=box];
    "Use contextual NLP\n(embeddings, LLMs)" [shape=box];
    "STOP: Not appropriate\nfor clinical use" [shape=box];

    "Corpus 500+\nwords?" -> "Primarily\nEnglish?" [label="yes"];
    "Corpus 500+\nwords?" -> "Report insufficient\ndata" [label="no"];
    "Primarily\nEnglish?" -> "Need semantic\ncontext?" [label="yes"];
    "Primarily\nEnglish?" -> "Flag language\nlimitation" [label="no"];
    "Need semantic\ncontext?" -> "Clinical/diagnostic\npurpose?" [label="no, dimension\npercentages sufficient"];
    "Need semantic\ncontext?" -> "Use contextual NLP\n(embeddings, LLMs)" [label="yes"];
    "Clinical/diagnostic\npurpose?" -> "STOP: Not appropriate\nfor clinical use" [label="yes"];
    "Clinical/diagnostic\npurpose?" -> "Proceed with\nLIWC analysis" [label="no"];
}

Quick Reference

LIWC-22 Major Dimension Hierarchy

Summary Variables (LIWC-22 Composites)

Key Parameters

Workflow

Copy this checklist and track progress:

LIWC Psycholinguistic Analysis Progress:
- [ ] Step 1: Validate corpus suitability and select tool
- [ ] Step 2: Preprocess text for dictionary-based analysis
- [ ] Step 3: Compute dimension percentages across all categories
- [ ] Step 4: Compute summary variables (Analytic, Clout, Authenticity, Tone)
- [ ] Step 5: Compare against reference baselines
- [ ] Step 6: Identify dominant dimensions and register profile
- [ ] Step 7: Distinguish process words from content words
- [ ] Step 8: Map dimensions to style replication priorities
- [ ] Step 9: Write findings to docs/analysis/16-liwc-psycholinguistic.md

Step 1: Validate Corpus Suitability and Select Tool

Suitability checks:

Tool selection:

digraph tool_selection {
    rankdir=TB;
    "Have LIWC-22\nlicense?" [shape=diamond];
    "Need negation\nhandling?" [shape=diamond];
    "Need custom\ncategories?" [shape=diamond];
    "Use LIWC-22" [shape=box];
    "Use SEANCE" [shape=box];
    "Use Empath" [shape=box];
    "Use custom\ndictionary approach" [shape=box];

    "Have LIWC-22\nlicense?" -> "Use LIWC-22" [label="yes"];
    "Have LIWC-22\nlicense?" -> "Need negation\nhandling?" [label="no"];
    "Need negation\nhandling?" -> "Use SEANCE" [label="yes"];
    "Need negation\nhandling?" -> "Need custom\ncategories?" [label="no"];
    "Need custom\ncategories?" -> "Use Empath" [label="yes"];
    "Need custom\ncategories?" -> "Use custom\ndictionary approach" [label="no, standard\ncategories sufficient"];
}

If corpus fails suitability: Report the failure in the output document. State what was checked, what failed, and why psycholinguistic dimension analysis is unreliable for this corpus.

Step 2: Preprocess Text for Dictionary-Based Analysis

Dictionary-based analysis has DIFFERENT preprocessing requirements than semantic NLP.

DO:

• Lowercase text for dictionary matching (LIWC and alternatives are case-insensitive lookups)
• Preserve contractions ("don't", "can't") -- many dictionaries include them as entries
• Expand common abbreviations only if the dictionary does not include them
• Remove URLs, HTML tags, user mentions/handles (these are not vocabulary)
• Normalize whitespace
• Keep punctuation separate from words (tokenize properly)

DO NOT:

• Stem or lemmatize (dictionaries contain specific word forms; stemming creates false matches and misses valid entries)
• Remove stopwords (function words ARE psycholinguistic data -- pronouns, articles, prepositions are primary carriers of process-level information)
• Remove numbers unless they dominate the corpus (LIWC has a "number" category)

import re
from collections import Counter

def preprocess_for_dictionary(text):
    """Clean text for LIWC-style dictionary matching.
    DO NOT stem. DO NOT remove stopwords. Function words are data."""
    text = re.sub(r'https?://\S+|www\.\S+', '', text)  # Remove URLs
    text = re.sub(r'@\w+', '', text)                     # Remove @mentions
    text = re.sub(r'<[^>]+>', '', text)                   # Remove HTML tags
    text = re.sub(r'&\w+;', '', text)                     # Remove HTML entities
    text = re.sub(r'\s+', ' ', text).strip()              # Normalize whitespace
    return text

def tokenize_for_dictionary(text):
    """Tokenize preserving contractions and word forms.
    Returns lowercase tokens for dictionary lookup."""
    text = preprocess_for_dictionary(text)
    # Split on whitespace and punctuation boundaries, keeping contractions intact
    tokens = re.findall(r"\b[\w']+\b", text.lower())
    return tokens

Step 3: Compute Dimension Percentages Across All Categories

For each LIWC-style dimension, count the number of tokens matching that dimension's dictionary and divide by total token count. Report percentages, not raw counts.

# Example using Empath (open-source LIWC alternative)
# pip install empath
from empath import Empath

lexicon = Empath()

def compute_dimension_percentages_empath(text):
    """Compute LIWC-style dimension percentages using Empath.
    Returns normalized percentages (each category as % of total words)."""
    result = lexicon.analyze(text, normalize=True)
    # Empath normalize=True divides by total category hits, not total words.
    # For LIWC-style percentages, we need per-word normalization.
    tokens = tokenize_for_dictionary(text)
    total_words = len(tokens)
    if total_words == 0:
        return None, 0

    # Re-analyze with raw counts for proper normalization
    raw_counts = lexicon.analyze(text, normalize=False)
    percentages = {
        cat: (count / total_words * 100) if count else 0.0
        for cat, count in raw_counts.items()
    }
    return percentages, total_words

# Example using a custom dictionary approach
# (When neither LIWC-22 nor Empath is available)
PSYCH_DIMENSIONS = {
    'cognitive_processes': {
        'insight': ['think', 'know', 'consider', 'understand', 'realize',
                     'recognize', 'believe', 'wonder', 'figure', 'idea'],
        'causation': ['because', 'cause', 'effect', 'hence', 'therefore',
                       'reason', 'since', 'result', 'lead', 'produce'],
        'discrepancy': ['should', 'would', 'could', 'ought', 'need',
                         'want', 'wish', 'hope', 'expect', 'desire'],
        'tentative': ['maybe', 'perhaps', 'guess', 'seem', 'appear',
                       'might', 'possibly', 'probably', 'suppose', 'unsure'],
        'certainty': ['always', 'never', 'definitely', 'certain', 'sure',
                       'absolute', 'clearly', 'obvious', 'undoubtedly', 'truly'],
        'differentiation': ['but', 'except', 'without', 'rather', 'instead',
                             'however', 'although', 'unless', 'unlike', 'whereas'],
    },
    'social_processes': {
        'social_general': ['talk', 'share', 'friend', 'together', 'group',
                            'join', 'meet', 'chat', 'community', 'people'],
        'family': ['mother', 'father', 'parent', 'family', 'brother',
                    'sister', 'daughter', 'son', 'wife', 'husband'],
        'affiliation': ['ally', 'bond', 'connect', 'belong', 'team',
                         'cooperate', 'help', 'support', 'trust', 'loyal'],
    },
    'affective_processes': {
        'positive_emotion': ['happy', 'love', 'joy', 'wonderful', 'great',
                              'excellent', 'good', 'nice', 'enjoy', 'pleased'],
        'negative_emotion': ['bad', 'wrong', 'terrible', 'awful', 'horrible',
                              'ugly', 'nasty', 'poor', 'miserable', 'pathetic'],
        'anxiety': ['worried', 'nervous', 'afraid', 'scared', 'anxious',
                     'fear', 'panic', 'tense', 'dread', 'uneasy'],
        'anger': ['hate', 'angry', 'furious', 'annoyed', 'irritated',
                   'rage', 'hostile', 'bitter', 'resentful', 'mad'],
        'sadness': ['sad', 'cry', 'grief', 'lonely', 'depressed',
                     'sorry', 'miss', 'hurt', 'mourn', 'heartbreak'],
    },
    'drives': {
        'achievement': ['win', 'success', 'earn', 'complete', 'accomplish',
                         'goal', 'achieve', 'master', 'excel', 'best'],
        'power': ['superior', 'control', 'command', 'dominate', 'authority',
                   'lead', 'influence', 'rule', 'force', 'power'],
        'reward': ['benefit', 'prize', 'gain', 'profit', 'reward',
                    'bonus', 'gift', 'valuable', 'worth', 'treasure'],
        'risk': ['danger', 'risk', 'threat', 'hazard', 'warning',
                  'caution', 'beware', 'vulnerable', 'unsafe', 'gamble'],
    },
    'biological_processes': {
        'body': ['body', 'hand', 'head', 'face', 'heart',
                  'skin', 'blood', 'arm', 'eye', 'brain'],
        'health': ['sick', 'pain', 'hospital', 'doctor', 'medicine',
                    'disease', 'cure', 'symptom', 'healthy', 'injury'],
        'ingestion': ['eat', 'drink', 'food', 'meal', 'hungry',
                       'taste', 'cook', 'dinner', 'breakfast', 'lunch'],
    },
    'perceptual_processes': {
        'see': ['see', 'look', 'watch', 'view', 'notice',
                 'observe', 'stare', 'glance', 'visible', 'bright'],
        'hear': ['hear', 'listen', 'sound', 'loud', 'quiet',
                  'noise', 'ring', 'voice', 'music', 'silent'],
        'feel': ['feel', 'touch', 'warm', 'cold', 'soft',
                  'hard', 'rough', 'smooth', 'sharp', 'heavy'],
    },
}

def compute_dimension_percentages_custom(tokens, dimensions=PSYCH_DIMENSIONS):
    """Compute dimension percentages using a custom dictionary.
    Returns nested dict: {dimension: {subcategory: percentage}}."""
    total_words = len(tokens)
    if total_words == 0:
        return None

    token_set = Counter(tokens)
    results = {}

    for dimension, subcategories in dimensions.items():
        dim_total = 0
        sub_results = {}
        for subcat, word_list in subcategories.items():
            count = sum(token_set.get(w, 0) for w in word_list)
            pct = count / total_words * 100
            sub_results[subcat] = {'count': count, 'percentage': round(pct, 3)}
            dim_total += count
        dim_pct = dim_total / total_words * 100
        results[dimension] = {
            'total_percentage': round(dim_pct, 3),
            'total_count': dim_total,
            'subcategories': sub_results,
        }

    # Coverage: what percentage of tokens matched any category
    all_matched = sum(d['total_count'] for d in results.values())
    results['_meta'] = {
        'total_words': total_words,
        'matched_words': all_matched,
        'coverage_pct': round(all_matched / total_words * 100, 2),
        'unmatched_pct': round((1 - all_matched / total_words) * 100, 2),
    }
    return results

Critical: percentages, not raw counts. Always report count / total_words * 100. Raw counts are uninterpretable without knowing corpus size. A corpus with 50 anger words could be intensely angry (50/500 = 10%) or emotionally bland (50/50,000 = 0.1%).

Step 4: Compute Summary Variables

If using LIWC-22, summary variables (Analytic, Clout, Authenticity, Emotional Tone) are computed automatically. For open-source tools, approximate these composites from available dimensions.

def compute_summary_approximations(dimension_results, tokens):
    """Approximate LIWC-22 summary variables from dimension percentages.
    These are APPROXIMATIONS, not validated equivalents of LIWC-22 composites."""
    total = len(tokens)
    if total == 0:
        return None

    cog = dimension_results.get('cognitive_processes', {}).get('total_percentage', 0)
    social = dimension_results.get('social_processes', {}).get('total_percentage', 0)
    affect = dimension_results.get('affective_processes', {}).get('total_percentage', 0)

    # Count function word categories
    articles = sum(1 for t in tokens if t in {'a', 'an', 'the'})
    prepositions = sum(1 for t in tokens if t in {
        'to', 'of', 'in', 'for', 'on', 'with', 'at', 'by', 'from',
        'about', 'into', 'through', 'during', 'before', 'after'
    })
    i_pronouns = sum(1 for t in tokens if t in {'i', 'me', 'my', 'mine', 'myself'})
    we_pronouns = sum(1 for t in tokens if t in {'we', 'us', 'our', 'ours', 'ourselves'})

    articles_pct = articles / total * 100
    prepositions_pct = prepositions / total * 100
    i_pron_pct = i_pronouns / total * 100
    we_pron_pct = we_pronouns / total * 100

    pos_emo = dimension_results.get('affective_processes', {}).get(
        'subcategories', {}).get('positive_emotion', {}).get('percentage', 0)
    neg_emo = dimension_results.get('affective_processes', {}).get(
        'subcategories', {}).get('negative_emotion', {}).get('percentage', 0)

    # Analytic approximation: high articles + prepositions + cognitive = formal/analytic
    analytic_raw = articles_pct + prepositions_pct + cog
    # Clout approximation: high we-pronouns, low I-pronouns, high social
    clout_raw = we_pron_pct + social - i_pron_pct
    # Authenticity approximation: high I-pronouns, high negative emotion, low articles
    authenticity_raw = i_pron_pct + neg_emo - articles_pct
    # Tone approximation: positive emotion vs negative emotion
    tone_raw = pos_emo - neg_emo

    return {
        'analytic_approx': round(analytic_raw, 2),
        'clout_approx': round(clout_raw, 2),
        'authenticity_approx': round(authenticity_raw, 2),
        'tone_approx': round(tone_raw, 2),
        '_note': 'These are directional approximations, NOT validated LIWC-22 percentile scores.',
    }

Important: If not using LIWC-22, label all summary variables as "approximations" and note they are directional indicators, not psychometrically validated percentile scores.

Step 5: Compare Against Reference Baselines

Dimension percentages are meaningless without comparison. A cognitive process rate of 12% means nothing until you know the reference rate is 11% (essentially average) or 6% (substantially elevated).

Reference baselines from published LIWC norms (approximate ranges):

Baseline comparison procedure:

1. Compute the corpus's dimension percentage for each category
2. Compare against the reference range in the table above
3. Flag categories that fall outside the typical range as "elevated" or "reduced"
4. Compute a deviation score: (corpus_rate - midpoint_of_range) / range_width
5. Rank dimensions by absolute deviation to identify the most distinctive features

REFERENCE_BASELINES = {
    'cognitive_processes': (10.0, 14.0),
    'social_processes': (8.0, 12.0),
    'affective_processes': (5.0, 8.0),
    'positive_emotion': (3.0, 5.0),
    'negative_emotion': (1.5, 3.0),
    'drives': (6.0, 9.0),
    'biological_processes': (1.0, 3.0),
    'perceptual_processes': (2.0, 4.0),
}

def compare_to_baselines(dimension_results):
    """Compare corpus dimension percentages against reference baselines.
    Returns deviation analysis for each dimension."""
    comparisons = []
    for dim, (low, high) in REFERENCE_BASELINES.items():
        if dim in dimension_results:
            corpus_pct = dimension_results[dim].get('total_percentage', 0)
        else:
            # Check subcategories
            for parent_dim in dimension_results.values():
                if isinstance(parent_dim, dict) and 'subcategories' in parent_dim:
                    if dim in parent_dim['subcategories']:
                        corpus_pct = parent_dim['subcategories'][dim]['percentage']
                        break
            else:
                continue

        midpoint = (low + high) / 2
        range_width = high - low
        deviation = (corpus_pct - midpoint) / range_width if range_width > 0 else 0

        status = 'typical'
        if corpus_pct > high:
            status = 'elevated'
        elif corpus_pct < low:
            status = 'reduced'

        comparisons.append({
            'dimension': dim,
            'corpus_pct': corpus_pct,
            'baseline_range': f"{low}-{high}%",
            'status': status,
            'deviation': round(deviation, 2),
        })

    # Sort by absolute deviation (most distinctive first)
    comparisons.sort(key=lambda x: abs(x['deviation']), reverse=True)
    return comparisons

Step 6: Identify Dominant Dimensions and Register Profile

Rank all dimensions by their percentage weight and by their deviation from baseline. The dominant dimensions define the corpus's psycholinguistic register.

Register classification:

Multi-dimension dominance: Most corpora show 2-3 co-dominant dimensions. Report the top 3 by deviation from baseline and describe the composite register (e.g., "Analytical-reflective with social-relational secondary emphasis").

Step 7: Distinguish Process Words from Content Words

This is a critical differentiation for style replication. Process words (function words: pronouns, articles, prepositions, auxiliary verbs) reveal HOW someone writes. Content words (nouns, verbs, adjectives, adverbs carrying topical meaning) reveal WHAT they write about.

Why this matters for style replication:

• Process words are more stable across topics and contexts -- they reflect habitual language patterns
• Content words vary with topic -- they reflect what the author is discussing, not how they discuss it
• A style replication system should prioritize matching process-word patterns (which persist across topics) over content-word patterns (which change with subject matter)

FUNCTION_WORD_CATEGORIES = {
    'personal_pronouns': {'i', 'me', 'my', 'mine', 'myself', 'you', 'your',
                           'yours', 'he', 'him', 'his', 'she', 'her', 'hers',
                           'it', 'its', 'we', 'us', 'our', 'they', 'them', 'their'},
    'articles': {'a', 'an', 'the'},
    'prepositions': {'to', 'of', 'in', 'for', 'on', 'with', 'at', 'by', 'from',
                      'about', 'into', 'through', 'during', 'before', 'after',
                      'above', 'below', 'between', 'under', 'over'},
    'auxiliary_verbs': {'is', 'am', 'are', 'was', 'were', 'be', 'been', 'being',
                         'have', 'has', 'had', 'do', 'does', 'did', 'will',
                         'would', 'shall', 'should', 'may', 'might', 'can',
                         'could', 'must'},
    'conjunctions': {'and', 'but', 'or', 'nor', 'for', 'yet', 'so',
                      'because', 'although', 'while', 'if', 'when', 'then'},
    'negations': {'not', 'no', 'never', 'neither', 'nobody', 'nothing',
                   'nowhere', 'nor'},
}

def compute_process_content_split(tokens):
    """Compute the ratio of process (function) words to content words."""
    all_function_words = set()
    for category_words in FUNCTION_WORD_CATEGORIES.values():
        all_function_words.update(category_words)

    total = len(tokens)
    if total == 0:
        return None

    function_count = sum(1 for t in tokens if t in all_function_words)
    content_count = total - function_count

    # Per-category breakdown
    category_breakdown = {}
    for cat, words in FUNCTION_WORD_CATEGORIES.items():
        cat_count = sum(1 for t in tokens if t in words)
        category_breakdown[cat] = {
            'count': cat_count,
            'percentage': round(cat_count / total * 100, 2),
        }

    return {
        'function_word_pct': round(function_count / total * 100, 2),
        'content_word_pct': round(content_count / total * 100, 2),
        'function_content_ratio': round(function_count / max(content_count, 1), 3),
        'category_breakdown': category_breakdown,
        'style_stability_note': (
            'Function word patterns are more stable across topics and '
            'should be prioritized for style replication. Content word '
            'patterns reflect topic, not style.'
        ),
    }

Step 8: Map Dimensions to Style Replication Priorities

Translate the dimension profile into a ranked list of linguistic registers that a style replication system should prioritize.

Priority mapping rules:

1. Highest priority: Dimensions with the largest deviation from baseline (most distinctive features)
2. Second priority: The process/content word ratio and function word category distribution (most stable style markers)
3. Third priority: Summary variable approximations (overall writing posture)
4. Lowest priority: Minor categories near baseline (these do not differentiate the writing)

def map_to_replication_priorities(baseline_comparisons, process_content, summary_vars):
    """Generate ranked style replication priorities from psycholinguistic analysis."""
    priorities = []

    # Priority 1: Most distinctive dimensions (by deviation from baseline)
    for comp in baseline_comparisons:
        if abs(comp['deviation']) > 0.5:  # Outside typical range
            priorities.append({
                'priority': 'HIGH',
                'feature': comp['dimension'],
                'direction': comp['status'],
                'corpus_rate': f"{comp['corpus_pct']:.1f}%",
                'baseline': comp['baseline_range'],
                'replication_note': f"Style replication MUST {'amplify' if comp['status'] == 'elevated' else 'suppress'} {comp['dimension']} language to match this corpus.",
            })
        elif abs(comp['deviation']) > 0.25:
            priorities.append({
                'priority': 'MEDIUM',
                'feature': comp['dimension'],
                'direction': comp['status'],
                'corpus_rate': f"{comp['corpus_pct']:.1f}%",
                'baseline': comp['baseline_range'],
                'replication_note': f"Style replication should lean {'toward' if comp['status'] == 'elevated' else 'away from'} {comp['dimension']} language.",
            })

    # Priority 2: Process word signature
    if process_content:
        fw_pct = process_content['function_word_pct']
        priorities.append({
            'priority': 'HIGH',
            'feature': 'function_word_ratio',
            'direction': f"{fw_pct:.1f}% function words",
            'corpus_rate': f"{fw_pct:.1f}%",
            'baseline': '55-65%',
            'replication_note': 'Function word patterns are the most stable style markers. Match this ratio closely.',
        })

    # Priority 3: Summary posture
    if summary_vars and '_note' in summary_vars:
        for key in ['analytic_approx', 'clout_approx', 'authenticity_approx', 'tone_approx']:
            val = summary_vars.get(key, 0)
            if abs(val) > 2.0:  # Meaningfully non-zero
                priorities.append({
                    'priority': 'MEDIUM',
                    'feature': key.replace('_approx', ''),
                    'direction': 'high' if val > 0 else 'low',
                    'value': val,
                    'replication_note': f"Writing posture leans {'formal/analytical' if key == 'analytic_approx' and val > 0 else 'personal/narrative' if key == 'analytic_approx' else key.replace('_approx', '')}.",
                })

    # Sort by priority
    priority_order = {'HIGH': 0, 'MEDIUM': 1, 'LOW': 2}
    priorities.sort(key=lambda x: priority_order.get(x['priority'], 3))

    return priorities

Step 9: Write Report

Write all findings to docs/analysis/16-liwc-psycholinguistic.md. See the report template below.

Report Output Template

The final report MUST be written to docs/analysis/16-liwc-psycholinguistic.md with this structure:

# LIWC Psycholinguistic Analysis

## Methodology
- **Tool:** [LIWC-22 / Empath / SEANCE / Custom dictionary -- specify which]
- **Corpus:** [N words total, N documents/texts, date range if available, source description]
- **Preprocessing:** [dictionary-safe: lowercased for matching, preserved contractions, removed URLs/HTML/mentions, NO stemming, NO stopword removal]
- **Dictionary coverage:** [X% of tokens matched at least one category; Y% unmatched]
- **Baseline comparison:** [reference norms used -- LIWC-22 published norms / approximated from literature]

## Corpus Suitability Assessment
- **Word count:** [N] words ([sufficient / marginal / insufficient])
- **Language:** [English / mixed -- percentage]
- **Register diversity:** [single register / multiple registers -- list them]
- **Authorship:** [single / multiple / unknown]
- **Overall suitability:** [suitable / suitable with caveats / unsuitable]
- **Confidence level:** [high (3,000+) / moderate (1,000-3,000) / low (500-1,000)]

## Dimension Percentages

### Major Dimensions
| Dimension | Corpus % | Baseline Range | Status | Deviation |
|-----------|----------|----------------|--------|-----------|
| Cognitive Processes | [X.X%] | 10-14% | [typical/elevated/reduced] | [+/-X.XX] |
| Social Processes | [X.X%] | 8-12% | ... | ... |
| Affective Processes | [X.X%] | 5-8% | ... | ... |
| Drives | [X.X%] | 6-9% | ... | ... |
| Biological Processes | [X.X%] | 1-3% | ... | ... |
| Perceptual Processes | [X.X%] | 2-4% | ... | ... |

### Subcategory Detail
[Table breaking down each major dimension into its subcategories with percentages]

### Summary Variable Approximations
| Variable | Value | Interpretation |
|----------|-------|----------------|
| Analytic (approx) | [X.XX] | [formal/analytical vs personal/narrative] |
| Clout (approx) | [X.XX] | [authoritative vs tentative] |
| Authenticity (approx) | [X.XX] | [personal/vulnerable vs guarded/distanced] |
| Emotional Tone (approx) | [X.XX] | [positive vs negative vs neutral] |

[If using LIWC-22, report actual percentile scores instead of approximations]

## Process vs Content Word Analysis

| Category | Percentage | Count |
|----------|-----------|-------|
| Function words (total) | [X.X%] | [N] |
| Content words (total) | [X.X%] | [N] |
| Function/Content ratio | [X.XXX] | -- |
| Personal pronouns | [X.X%] | [N] |
| Articles | [X.X%] | [N] |
| Prepositions | [X.X%] | [N] |
| Auxiliary verbs | [X.X%] | [N] |
| Conjunctions | [X.X%] | [N] |
| Negations | [X.X%] | [N] |

**Style stability note:** [Interpretation of function word patterns as stable style markers vs content words as topic-dependent]

## Dominant Dimension Profile

### Register Classification
- **Primary register:** [Analytical-reflective / Social-relational / Expressive-positive / etc.]
- **Secondary register:** [if applicable]
- **Tertiary register:** [if applicable]

### Most Distinctive Features (ranked by deviation from baseline)
1. [Dimension]: [corpus %] vs [baseline range] -- [interpretation]
2. [Dimension]: ...
3. [Dimension]: ...

## Style Replication Priorities

| Priority | Feature | Direction | Corpus Rate | Baseline | Replication Guidance |
|----------|---------|-----------|-------------|----------|---------------------|
| HIGH | [feature] | [elevated/reduced] | [X.X%] | [range] | [specific guidance] |
| HIGH | function_word_ratio | [X.X%] | [X.X%] | 55-65% | [guidance] |
| MEDIUM | [feature] | ... | ... | ... | ... |
| ... | ... | ... | ... | ... | ... |

## Limitations and Caveats
- Dictionary-based categorization captures vocabulary composition, NOT semantic meaning. The word "cut" is categorized identically whether it refers to surgery, budget reductions, or film editing.
- No handling of negation: "not happy" is counted as positive emotion. [If using SEANCE, this limitation is partially mitigated.]
- No handling of sarcasm, irony, or figurative language.
- Polysemy (words with multiple meanings) inflates some categories and deflates others unpredictably.
- [Corpus-specific limitations from Step 1]
- [Confidence level acknowledgment and what it means for reliability]
- Summary variable approximations are directional indicators only, not psychometrically validated equivalents of LIWC-22 composites.
- These results should NOT be used for psychological diagnosis, clinical assessment, or evaluative judgments about mental health or character.
- Dimension percentages reflect vocabulary composition in THIS corpus in THIS context. The same author may produce different dimension profiles in different writing contexts.
- Results from open-source alternatives (Empath, SEANCE, custom dictionaries) may differ from LIWC-22 due to different dictionary compositions, category boundaries, and normalization methods.

## Cross-Tool Consistency Note
[If multiple tools were used, report where they agreed and diverged. If only one tool was used, note that cross-validation with a second tool would strengthen confidence.]

## References
- Boyd, R.L., Ashokkumar, A., Seraj, S., & Pennebaker, J.W. (2022). The development and psychometric properties of LIWC-22. University of Texas at Austin.
- Pennebaker, J.W., Boyd, R.L., Jordan, K., & Blackburn, K. (2015). The development and psychometric properties of LIWC2015. University of Texas at Austin.
- Tausczik, Y.R. & Pennebaker, J.W. (2010). The psychological meaning of words: LIWC and computerized text analysis methods. *Journal of Language and Social Psychology*, 29(1), 24-54.
- Fast, E., Chen, B., & Bernstein, M.S. (2016). Empath: Understanding topic signals in large-scale text. *CHI*.
- Crossley, S.A., Kyle, K., & McNamara, D.S. (2017). Sentiment analysis and social cognition engine (SEANCE). *Behavior Research Methods*, 49(3), 803-821.
- [Additional references as applicable]

Good Patterns

• Report percentages, never raw counts -- raw counts without corpus size normalization are meaningless
• Use validated dictionaries from published research (LIWC, Empath, SEANCE) rather than ad-hoc word lists
• Compare against reference baselines -- a dimension percentage in isolation has no interpretive value
• Analyze multiple dimensions simultaneously -- a corpus's psycholinguistic profile is the PATTERN across dimensions, not any single dimension
• Distinguish process words from content words -- function word patterns are stable style markers; content words reflect topic
• Preserve function words during preprocessing -- DO NOT remove stopwords; pronouns, articles, and prepositions ARE the psycholinguistic data
• Rank dimensions by baseline deviation -- the most distinctive dimensions (lar

…(truncated)